Niklas Haas: > On Fri, 28 Feb 2025 10:31:19 +0800 Zhao Zhili <quinkbl...@foxmail.com> wrote: >> Cc haasn. >> >> Libswscale in under refactor. Does current asm works after refactor, or they >> need to be refactored or >> rewrite after? If it’s the second case, maybe we should hold on to do more >> asm with libswscale >> before hassn work done. > > No, almost all current asm will be unused after the rewrite. There are some we > can in theory reuse, but for the most part, it doesn't seem to be worth it. > > Especially for the very bespoke functions like this one. > > For context, in general, the focus in nu-swscale is to focus more on smaller, > flexible primitives and have the calling code combine them as needed. So > instead > of a "brga_to_y" function, you would have a sequence that looks like this: > > Operation list: > [ u8 XXXX -> dddX] SWS_OP_READ : 4 elem(s) packed >> 0 > [ u8 ...X -> dddX] SWS_OP_SWIZZLE : 2103 > [ u8 ...X -> dddX] SWS_OP_CONVERT : u8 -> f32 > [f32 ...X -> .XXX] SWS_OP_LINEAR : dot3 [[0.299000 0.587000 0.114000 > 0 0] [0 1 0 0 0] [0 0 1 0 0] [0 0 0 1 0]] > [f32 .XXX -> .XXX] SWS_OP_DITHER : 16x16 {255 _ _ _} > [f32 .XXX -> dXXX] SWS_OP_CONVERT : f32 -> u8 > [ u8 .XXX -> XXXX] SWS_OP_WRITE : 1 elem(s) packed >> 0 > > Where each low-level implementation can combine one, or multiple, such > operations together. For example, in the current prototype, SWS_OP_CONVERT and > SWS_OP_WRITE can be fused together into a single implementation. > > Note also the conversion to float. I found that the cost of going through > floats seems to be lower on average, across all tested platforms, than the > extra cost of dealing with integers (which require extra shifting, extra > dithering, and extra width conversions - all of which exceed the cost of just > one extra float->int conversion step). This also comes with improved accuracy. > But what about bitexactness?
- Andreas _______________________________________________ ffmpeg-devel mailing list ffmpeg-devel@ffmpeg.org https://ffmpeg.org/mailman/listinfo/ffmpeg-devel To unsubscribe, visit link above, or email ffmpeg-devel-requ...@ffmpeg.org with subject "unsubscribe".